MiniMax releases M3 large model: programming ability surpasses GPT-5.5, supports native multimodal desktop control

CryptoWorld News reports that MiniMax officially released its large model M3 today.
M3 is currently the only open-source large model that fully integrates the three cutting-edge elements: programming, ultra-long context, and native multimodality, with plans to open-source the weights within 10 days.
It achieves international leading levels in code generation, intelligent agents, and desktop control, available through MiniMax code, token plan, and API.
M3 pioneers a sparse attention architecture called MSA, which aggregates hits in the KV blocks to query, making memory access four times faster than Flash-sparse-attention.
With a context length of around one million, the new architecture reduces per-token computation to one-twentieth of the previous generation, achieving 9x faster pre-filling and 15x faster decoding.
On SWE-bench pro, M3 scored 59.0%, surpassing GPT-5.5 and Gemini 3.1 pro, approaching Opus 4.7.
In the Hopper-optimized FP8 operator task, it autonomously invoked tools 1,959 times within 24 hours, increasing hardware utilization from 7.6% to 71.3%, a 9.4x acceleration.
The API is now live, offering inference and fast modes, with weights planned to be open-sourced within 10 days.
View Original
This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.
  • Reward
  • 8
  • 1
  • Share
Comment
Add a comment
Add a comment
ThereIsTvlInTheWind
· 5h ago
SWE-bench pro 59% surpasses GPT-5.5, really standing out in coding ability.
View OriginalReply0
SymbolsInTheReflection
· 5h ago
MSA architecture accesses memory 4 times faster, making Flash-sparse-attention a mere background.
View OriginalReply0
ThereAreCatsInTheContract.
· 5h ago
KV block aggregation hit query, detailed analysis of technical details and related papers
View OriginalReply0
PerpPulse
· 5h ago
Is Gemini 3.1 being surpassed? Google is feeling the pressure.
View OriginalReply0
TeaAndSlippage
· 5h ago
Programming + ultra-long context + native multimodal three-in-one integration—an unparalleled open-source path
View OriginalReply0
GateUser-f7b40cee
· 5h ago
MiniMax code and API are now available for testing; just go for it.
View OriginalReply0
DuskStop-LossLine
· 5h ago
Desktop control reaches international leading standards, AI Agent implementation takes another step forward
View OriginalReply0
AirdropMileCounter
· 5h ago
Pre-filling 9 times decoding for 15 times, this acceleration ratio is truly outrageous
View OriginalReply0
  • Pinned